science
i work on interpretable machine learning for biological data: models whose decisions can be read, checked and argued with. my master's research at utfpr (ppgtca) is on the soybean microbiome.
does the model learn the disease, or the study?
a cervicovaginal microbiome meta-analysis on cervical cancer, reprocessed from the raw 16s reads: 10 public projects, 889 samples after quality control. with random cross-validation the classifier reaches a macro-auc of 0.81. holding out one whole project at a time, it drops to 0.60. project of origin explains about 16 times more variation than clinical condition.
- method
- qiime 2, permanova, gllvm, random forest, leave-one-project-out
- published
- latin.science 2026 (latinoware)
soil microbiome under water stress
a full replication of hagen et al. (2024), rebuilt from the raw data: 623 samples of 16s sequencing, five differential-abundance methods, and a random forest explained with shap at five taxonomic ranks. the genus level did best, with an auc of 0.979, close to the original paper, and shap pointed to the same marker taxa.
- method
- qiime 2, dada2, phyloseq, random forest, shap
- published
- sbproppg 2026 (utfpr)
from the microbiome to a prescription map
a classifier reads water stress from soil bacteria, shap names the marker taxa, and the output becomes a prescription map: a dose of bio-input for each zone of the field. on a synthetic 25 ha plot, variable-rate application used about 52% less input than a uniform one. built on simulated data, to validate the pipeline end to end.
- method
- python, random forest, shap, nested cross-validation
computer vision at embrapa
from 2020 to 2021, as an undergraduate researcher: fruit detection in orchards with yolo and opencv, and automatic apple counting that saved 100+ hours of manual work per harvest.
published papers
- 2026validação por projeto em meta-análises de microbioma: um estudo de caso em câncer cervical · latin.science (latinoware)
- 2026validação multiprojeto em análise de dados físico-químicos do solo · latin.science (latinoware)
- 2026aprendizado de máquina interpretável e microbioma do solo: replicação de pipeline para detecção de estresse hídrico · sbproppg (utfpr)
- 2021best short paper award · xvii latinoware
- 2020–21two papers on computer vision applied to fruit growing · embrapa